[pt] O desenvolvimento de técnicas de aprendizado auto-supervisionado vem ganhando muita visibilidade na área de Visão Computacional pois possibilita o pré-treinamento de redes neurais profundas sem a necessidade de dados anotados. Em alguns domínios, as anotações são custosas, pois demandam muito trabalho especializado para a rotulação dos dados. Esse problema é muito comum no setor de Óleo e Gás, onde existe um vasto volume de dados não interpretados. O presente trabalho visa aplicar a técnica de aprendizado auto-supervisionado denominada Masked Autoencoders para pré-treinar modelos Vision Transformers com dados sísmicos. Para avaliar o pré-treino, foi aplicada a técnica de transfer learning para o problema de segmentação de fácies sísmicas. Na fase de pré-treinamento foram empregados quatro volumes sísmicos distintos. Já para a segmentação foi utilizado o dataset Facies-Mark e escolhido o modelo da literatura Segmentation Transformers. Para avaliação e comparação da performance da metodologia foram empregadas as métricas de segmentação utilizadas pelo trabalho de benchmarking de ALAUDAH (2019). As métricas obtidas no presente trabalho mostraram um resultado superior. Para a métrica frequency weighted intersection over union, por exemplo, obtivemos um ganho de 7.45 por cento em relação ao trabalho de referência. Os resultados indicam que a metodologia é promissora para melhorias de problemas de visão computacional em dados sísmicos. / [en] The development of self-supervised learning techniques has gained a lot of visibility in the field of Computer Vision as it allows the pre-training of deep neural networks without the need for annotated data. In some domains, annotations are costly, as they require a lot of specialized work to label the data. This problem is very common in the Oil and Gas sector, where there is a vast amount of uninterpreted data. The present work aims to apply the self-supervised learning technique called Masked Autoencoders to pre-train Vision Transformers models with seismic data. To evaluate the pre-training, transfer learning was applied to the seismic facies segmentation problem. In the pre-training phase, four different seismic volumes were used. For the segmentation, the Facies-Mark dataset was used and the Segmentation Transformers model was chosen from the literature. To evaluate and compare the performance of the methodology, the segmentation metrics used by the benchmarking work of ALAUDAH (2019) were used. The metrics obtained in the present work showed a superior result. For the frequency weighted intersection over union (FWIU) metric, for example, we obtained a gain of 7.45 percent in relation to the reference work. The results indicate that the methodology is promising for improving computer vision problems in seismic data.

Nanopore sequencing is the next generation ofsequencing methods which promises to deliver cheaper andmore portable genome sequencing capabilities. A single DNAor RNA strand is passed through a nanopore nested in anartificial membrane with an electric potential applied across it.The nucleotide bases of the helix then interact with the ioniccurrent in the nanopore, resulting in a unique signal that canbe translated into the correct corresponding nucleotide sequence.This project investigated whether features of the raw signal datacould be used as predictive indicators of the duration time ofeach nucleotide base in the nanopore. This is done in orderto segment the signal before translation. The training data setused came from the sequenced DNA molecules of an E. Colibacterium. Distribution candidates were fitted to a histogram ofthe duration data of the training set. Features of the currentsignal and distribution parameters were correlated in orderinvestigate if a linear predictive model could be created. Theresults indicate that the feature zero-crossings is not an optimaloption for construction of a linear model, while the large jumpsand moving variance features often generate linear patterns. The parameter of the Log-logistic distribution had the best fit withthe lowest relative root mean square deviation (rRMSD) of 2.7%. / Nanopore sequencing är nästa generationensmetod för DNA sekvensering som kommer att bidra medbilligare och mer portabla sekvenseringsmöjligheter. Metodeninnebär att en enkelsträngad DNA eller RNA molekyl passerargenom porer i nanostorlek, placerade i ett artificiellt membransamtidigt som en elektrisk potential appliceras över membranet.Nukleotiderna i genmolekylen interagerar med jonströmmen iporen, vilket resulterar i en unik signal som kan översättas tillden korresponderande sekvensen av nukleotider som passerat.Detta projekt gick ut på att undersöka om egenskaper frånsignalen kan användas som predikativa indikatorer för varaktighetensom varje nukleotid befinner sig i membranporen. Dettaför att sedan kunna segmentera signalen före översättningen tillDNA sekvensen. Träningsdata som användes är sekvenserad DNAfrån en E. Coli bakterie. Kandiderande sannolikhetsfördelningaranpassades till ett histogram som beskriver varaktigheten.Egenskaperna och parametrar från fördelningarna korreleradesför att skapa en linjär modell. Resultatet visade att antaletskärningar i x-axeln som signalegenskap inte är det optimalavalet för konstruktion av en linjär modell. Skillnaden mellan två signalvärden som är mindre än en varierbar konstant ochglidande variansen som signalegenskaper genererar ofta linjäramönster. Resultatet visade även att sannolikhetsfördelningen Loglogistichade lägst relativ medelkvadratavvikelse (rRMSD) på 2.7%. / Kandidatexjobb i elektroteknik 2021, KTH, Stockholm

Representation Learning is ubiquitous in state-of-the-art machine learning workflow, including data exploration/visualization, data preprocessing, data model learning, and model interpretations. However, the majority of the newly proposed Representation Learning methods are more suitable for problems with a large amount of data. Applying these methods to problems with a limited amount of data may lead to unsatisfactory performance. Therefore, there is a need for developing Representation Learning methods which are tailored for problems with ``small data", such as, clinical and biomedical data analytics. In this dissertation, we describe our studies of tackling the challenging clinical and biomedical data analytics problem from four perspectives: data preprocessing, temporal data representation learning, output representation learning, and joint input-output representation learning. Data scaling is an important component in data preprocessing. The objective in data scaling is to scale/transform the raw features into reasonable ranges such that each feature of an instance will be equally exploited by the machine learning model. For example, in a credit flaw detection task, a machine learning model may utilize a person's credit score and annual income as features, but because the ranges of these two features are different, a machine learning model may consider one more heavily than another. In this dissertation, I thoroughly introduce the problem in data scaling and describe an approach for data scaling which can intrinsically handle the outlier problem and lead to better model prediction performance. Learning new representations for data in the unstandardized form is a common task in data analytics and data science applications. Usually, data come in a tubular form, namely, the data is represented by a table in which each row is a feature (row) vector of an instance. However, it is also common that the data are not in this form; for example, texts, images, and video/audio records. In this dissertation, I describe the challenge of analyzing imperfect multivariate time series data in healthcare and biomedical research and show that the proposed method can learn a powerful representation to encounter various imperfections and lead to an improvement of prediction performance. Learning output representations is a new aspect of Representation Learning, and its applications have shown promising results in complex tasks, including computer vision and recommendation systems. The main objective of an output representation algorithm is to explore the relationship among the target variables, such that a prediction model can efficiently exploit the similarities and potentially improve prediction performance. In this dissertation, I describe a learning framework which incorporates output representation learning to time-to-event estimation. Particularly, the approach learns the model parameters and time vectors simultaneously. Experimental results do not only show the effectiveness of this approach but also show the interpretability of this approach from the visualizations of the time vectors in 2-D space. Learning the input (feature) representation, output representation, and predictive modeling are closely related to each other. Therefore, it is a very natural extension of the state-of-the-art by considering them together in a joint framework. In this dissertation, I describe a large-margin ranking-based learning framework for time-to-event estimation with joint input embedding learning, output embedding learning, and model parameter learning. In the framework, I cast the functional learning problem to a kernel learning problem, and by adopting the theories in Multiple Kernel Learning, I propose an efficient optimization algorithm. Empirical results also show its effectiveness on several benchmark datasets. / Computer and Information Science

<p dir="ltr">As cameras increasingly populate urban environments for surveillance, the threat of data breaches and losses escalates as well. The rapid advancements in generative Artificial Intelligence have greatly simplified the replication of individuals’ appearances from video footage. This capability poses a grave risk as malicious entities can exploit it for various nefarious purposes, including identity theft and tracking individuals’ daily activities to facilitate theft or burglary.</p><p dir="ltr">To reduce reliance on video surveillance systems, this study introduces Urban Seismic Event Detection (USED), a deep learning-based technique aimed at extracting information about urban seismic events. Our approach involves synthesizing training data through a small batch of manually labelled field data. Additionally, we explore the utilization of unlabeled field data in training through semi-supervised learning, with the implementation of a mean-teacher approach. We also introduce pre-processing and post-processing techniques tailored to seismic data. Subsequently, we evaluate the trained models using synthetic, real, and unlabeled data and compare the results with recent statistical methods. Finally, we discuss the insights gained and the limitations encountered in our approach, while also proposing potential avenues for future research.</p>

Satellite remote sensing has revolutionized the acquisition of large amounts of data, employing both active and passive sensors to capture critical information about our planet. These data can be analysed by using deep learning methodologies that demonstrate excellent capabilities in extracting the semantics from the data. However, one of the main challenges in exploiting the power of deep learning for remote sensing applications is the lack of labeled training data. Deep learning architectures, typically demand substantial quantities of training samples to achieve optimal performance. Motivated by the above-mentioned challenges, this thesis focuses on the limited availability of labeled datasets. These challenges include issues such as ambiguous labels in case of large-scale remote sensing datasets, particularly when dealing with the analysis of multi-source satellite remote sensing images. By employing novel deep learning techniques and cutting-edge methodologies, this thesis endeavors to contribute to advancements in the field of remote sensing. In this thesis, the problems related to limited labels are solved in several ways by developing (i) a novel spectral index generative adversarial network to augment real training samples for generating class-specific remote sensing data to provide a large number of labeled samples to train a neural-network classifier; (ii) a mono- and dual-regulated contractive-expansive-contractive convolutional neural network architecture to incorporate spatial-spectral information of multispectral data and minimize the loss in the feature maps and extends this approach to the analysis of hyperspectral images; (iii) a hybrid deep learning architecture with a discrete wavelet transform and attention mechanism to deal with few labeled samples for scene-based classification of multispectral images; and (iv) a weak supervised semantic learning technique that utilises weak or low-resolution labeled samples with multisource remote sensing images for predicting pixel-wise land-use-land-cover maps. The experiments show that the proposed approaches perform better than the state-of-the-art methods on different benchmark datasets and in different conditions.

Identification of travelers' transportation modes is a fundamental step for various problems that arise in the domain of transportation such as travel demand analysis, transport planning, and traffic management. This thesis aims to identify travelers' transportation modes purely based on their GPS trajectories. First, a segmentation process is developed to partition a user's trip into GPS segments with only one transportation mode. A majority of studies have proposed mode inference models based on hand-crafted features, which might be vulnerable to traffic and environmental conditions. Furthermore, the classification task in almost all models have been performed in a supervised fashion while a large amount of unlabeled GPS trajectories has remained unused. Accordingly, a deep SEmi-Supervised Convolutional Autoencoder (SECA) architecture is proposed to not only automatically extract relevant features from GPS segments but also exploit useful information in unlabeled data. The SECA integrates a convolutional-deconvolutional autoencoder and a convolutional neural network into a unified framework to concurrently perform supervised and unsupervised learning. The two components are simultaneously trained using both labeled and unlabeled GPS segments, which have already been converted into an efficient representation for the convolutional operation. An optimum schedule for varying the balancing parameters between reconstruction and classification errors are also implemented. The performance of the proposed SECA model, trip segmentation, the method for converting a raw trajectory into a new representation, the hyperparameter schedule, and the model configuration are evaluated by comparing to several baselines and alternatives for various amounts of labeled and unlabeled data. The experimental results demonstrate the superiority of the proposed model over the state-of-the-art semi-supervised and supervised methods with respect to metrics such as accuracy and F-measure. / Master of Science / Identifying users' transportation modes (e.g., bike, bus, train, and car) is a key step towards many transportation related problems including (but not limited to) transport planning, transit demand analysis, auto ownership, and transportation emissions analysis. Traditionally, the information for analyzing travelers' behavior for choosing transport mode(s) was obtained through travel surveys. High cost, low-response rate, time-consuming manual data collection, and misreporting are the main demerits of the survey-based approaches. With the rapid growth of ubiquitous GPS-enabled devices (e.g., smartphones), a constant stream of users' trajectory data can be recorded. A user's GPS trajectory is a sequence of GPS points, recorded by means of a GPS-enabled device, in which a GPS point contains the information of the device geographic location at a particular moment. In this research, users' GPS trajectories, rather than traditional resources, are harnessed to predict their transportation mode by means of statistical models. With respect to the statistical models, a wide range of studies have developed travel mode detection models using on hand-designed attributes and classical learning techniques. Nonetheless, hand-crafted features cause some main shortcomings including vulnerability to traffic uncertainties and biased engineering justification in generating effective features. A potential solution to address these issues is by leveraging deep learning frameworks that are capable of capturing abstract features from the raw input in an automated fashion. Thus, in this thesis, deep learning architectures are exploited in order to identify transport modes based on only raw GPS tracks. It is worth noting that a significant portion of trajectories in GPS data might not be annotated by a transport mode and the acquisition of labeled data is a more expensive and labor-intensive task in comparison with collecting unlabeled data. Thus, utilizing the unlabeled GPS trajectory (i.e., the GPS trajectories that have not been annotated by a transport mode) is a cost-effective approach for improving the prediction quality of the travel mode detection model. Therefore, the unlabeled GPS data are also leveraged by developing a novel deep-learning architecture that is capable of extracting information from both labeled and unlabeled data. The experimental results demonstrate the superiority of the proposed models over the state-of-the-art methods in literature with respect to several performance metrics.

Accurate delineation of anatomical structures in the white matter of the human brain is of paramount importance for multiple applications, such as neurosurgical planning, characterization of neurological disorders, and connectomic studies. Diffusion Magnetic Resonance Imaging (dMRI) techniques can provide, in-vivo, a mathematical representation of thousands of fibers composing such anatomical structures, in the form of 3D polylines called streamlines. Given this representation, a task of invaluable interest is known as white matter bundle segmentation, whose aim is to virtually group together streamlines sharing a similar pathway into anatomically meaningful structures, called white matter bundles. Obtaining a good and reliable bundle segmentation is however not trivial, mainly because of the intrinsic complexity of the data. Most of the current methods for bundle segmentation require extensive neuroanatomical knowledge, are time consuming, or are not able to adapt to different data settings. To overcome these limitations, the main goal of this thesis is to develop a new automatic method for accurate white matter bundle segmentation, by exploiting, combining and extending multiple up-to-date supervised learning techniques. The main contribution of the project is the development of a novel streamline-based bundle segmentation method based on binary linear classification, which simultaneously combines information from atlases, bundle geometries, and connectivity patterns. We prove that the proposed method reaches unprecedented quality of segmentation, and that is robust to a multitude of diverse settings, such as when there are differences in bundle size, tracking algorithm, and/or quality of dMRI data. In addition, we show that some of the state-of-the-art bundle segmentation methods are deeply affected by a geometrical property of the shape of the bundles to be segmented, their fractal dimension. Important factors involved in the task of streamline classification are: (i) the need for an effective streamline distance function and (ii) the definition of a proper feature space. To this end, we compare some of the most common streamline distance functions available in the literature and we provide some guidelines on their practical use for the task of supervised bundle segmentation. Moreover, we investigate the possibility to include, in a streamline-based segmentation method, additional information to the typically employed streamline distance measure. Specifically, we provide evidence that considering additional anatomical information regarding the cortical terminations of the streamlines and their proximity to specific Regions of Interest (ROIs) helps to improve the results of bundle segmentation. Lastly, significant attention is paid to reproducibility in neuroscience. Following the FAIR (Findable, Accessible, Interoperable and Reusable) Data Principles, we have integrated our pipelines of analysis into an online open platform devoted to promoting reproducibility of scientific results and to facilitating knowledge discovery.

DeepPaint is a semi-automated segmentation tool that utilises a U-net architecture to performbinary segmentation. To maximise the model’s performance and minimise user time, it isadvisable to apply Transfer Learning (TL) and reuse a model trained on a similar segmentationtask. However, due to the sensitivity of medical data and the unique properties of certainsegmentation tasks, TL is not feasible for some applications. In such circumstances, SelfSupervised Learning (SSL) emerges as the most viable option to minimise the time spent inDeepPaint by a user. Various pretext tasks, exploring both corruption segmentation and corruption restoration, usingsuperpixels and square patches, were designed and evaluated. With a limited number ofiterations in both the pretext and downstream tasks, significant improvements across fourdifferent datasets were observed. The results reveal that SSL models, particularly those pretrained on corruption segmentation tasks where square patches were corrupted, consistentlyoutperformed models without pre-training, with regards to a cumulative Dice SimilarityCoefficient (DSC). To examine whether a model could learn relevant features from a pretext task, Centred KernelAlignment (CKA) was used to measure the similarity of feature spaces across a model's layersbefore and after fine-tuning on the downstream task. Surprisingly, no significant positivecorrelation between downstream DSC and CKA was observed in the encoder, likely due to thelimited fine-tuning allowed. Furthermore, it was examined whether pre-training on the entiredataset, as opposed to only the training subset, yielded different downstream results. Asexpected, significantly higher DSC in the downstream task is more likely if the model hadaccess to all data during the pretext task. The differences in downstream segmentationperformance between models that accessed different data subsets during pre-training variedacross datasets.

This study addresses the challenge of identifying anomalies within multivariate time series data, focusing specifically on the operational parameters of gas turbine combustion systems. In search of an effective detection method, the research explores the application of three distinct machine learning methods: the Long Short-Term Memory (LSTM) autoencoder, the Self-Organizing Map (SOM), and the Density-Based Spatial Clustering of Applications with Noise (DBSCAN). Through the experiment, these models are evaluated to determine their efficacy in anomaly detection. The findings show that the LSTM autoencoder not only surpasses its counterparts in performance metrics but also shows a unique capability to identify the underlying causes of detected anomalies. This paper delves into the comparative analysis of these techniques and discusses the implications of the models in maintaining the reliability and safety of gas turbine operations.

This thesis examines the effectiveness of Graph Neural Networks (GNNs) in detecting money laundering activities using transaction data with unreliable labels. It analyses how weakly supervised learning, specifically with GNNs, manages the challenges posed by incomplete and inaccurate labels in anti-money laundering (AML) detection. The thesis utilizes simulated transaction data to compare the performance of GNNs against statistical models. This was done by generating various datasets with the AMLSim tool, and evaluating the node classification performance of different statistical machine learning models and GNNs. The findings indicate that GNNs, due to their ability to find relationships in graph structures, demonstrate superior performance in scenarios with incomplete and inaccurate labels. The findings also indicate that inaccurate positive labels has a great negative effect on the performance, showing the label importance of money launderers in graph data. This research provides possible improvements for anti-money laundering detection by employing GNNs to manage challenges in real-world data.